version: "1.0.0" name: open-doors-and-drawers description: Open a specified hinged, sliding, or pull-out closure using robot perception and manipulation, including room doors, cabinet doors, drawers, microwave doors, refrigerator doors, and balcony or glass sliding doors. Use when the robot must open a closed or partially open door or drawer to gain access, expose its contents, or pass through. Do not use for closing a door or drawer, or when the target is already sufficiently open for the requested task.

Open doors and drawers

<!-- codex-coordinate-contract:start --> The latest image in the conversation represents the current observed state. Older images are historical context; their points, bounding boxes, and object-part locations do not describe the current view. After any motion, re-locate the target in the latest image; do not shift, rotate, or scale an older location estimate to choose a new click.

Before EVERY point selection, inspect the latest attached clickable head image and re-locate the physical target. Use the full attached image as a relative coordinate canvas: integer u (column) and v (row) in 0..1000, with top-left (0,0) and bottom-right (1000,1000). Do not use encoded image pixels or crop coordinates. Bind every point, including every row of a multi-point call, to that exact current image_id. Never copy locations from older images, earlier reasoning, reference photos, tracking text, or overlay labels. Wrist images are observation-only and cannot be clicked. If the point is rejected, obtain a fresh head image and ground again; do not guess an image_id. <!-- codex-coordinate-contract:end -->

<!-- codex-task-lifecycle:start --> Follow this procedure only while open-doors-and-drawers is the active task Skill loaded by activate_skill. Its presence in history, including after compaction, does not make it active. After success and all required post-actions, return to the parent task; activate another matching Skill only when its procedure is needed. Instructions below to stay, retry, or exit describe pursuit of success and do not prohibit a change of approach or handoff to the parent task. <!-- codex-task-lifecycle:end -->

Use the currently advertised behavior_v2 MCP tools and follow the baseline Skill. This Skill opens the requested closure; it must never command a closing motion. Successful completion requires a hinged door opened fully, as close to 90 degrees as the swing allows; or a long low TV stand drawer opened by TV stand drawer special case. Stopping a current rotation, bind, plan, or pull is a sub-step change, not Skill exit and not session end.

Entry and observations

• Enter only when the requested door, cabinet door, drawer, or sliding closure

is closed or not yet open far enough for the parent task. If an already-open door occupies the approach to the next target, first complete Clear around an opened door before the next, then continue with staging and reach.

• Use a current head-camera image to identify the requested closure, its handle,

and its mechanism. Reuse a model-visible post-action head image when it is current and usable. Call capture_head_camera only when no such image is available or the relevant geometry is ambiguous.

• For a hinged closure, visually determine whether its hinge is on the left or

right before grasp planning. Do not infer the hinge side from memory or from the handle side alone. If the hinge cannot be identified, obtain a better observation instead of guessing.

• A drawer or sliding door has no hinge side. Never invent one. A long low

TV stand drawer uses only TV stand drawer special case after reach. It must not enter Plan and grasp the handle, Retract before a new grasp plan, Probe the grasp and open the hinged door, or Clear around an opened door before the next. Those four sections are the hinged-door path. The numeric pull loop is only for hinged closures.

Establish a frontal staging pose

Establish a safe frontal staging pose before reaching for the handle.

From each current head image used for approach, state the facing judgment in words: facing, or not facing, plus a short note about the panel plane. Judge facing from the movable panel's plane. Treat the movable panel like a poster on a wall. Facing means the top edge and the bottom edge of that panel look level (horizontal). Do not judge facing from the left or right side edges, their lengths, or their depths. A large visible top surface after lowering the trunk does not mean not facing.

Not facing means the top edge or the bottom edge is clearly tilted. Visible handles, or the cabinet sitting in the middle of the image, help find the panel; facing is the panel-plane look. If that evidence is mixed, say not facing.

move_to_reach_point keeps the current chassis yaw and approaches the selected point in a straight line, so the panel's orientation stays whatever it already is. It does not weave around furniture. Any obstacle on that line will stall or jam the robot. Call it only from a pose whose straight corridor to the selected point is clear.

When the stated judgment is facing and the robot is already safely in front of the panel, continue to Set trunk height after facing, then Move into reach.

When the stated judgment is not facing, finish this section first. When the panel is still seen at a large inclination, first get farther from the panel so there is room to turn, then rotate until the panel plane faces the robot, then state the facing judgment again from the latest image. After that later judgment is facing, continue to Set trunk height after facing, then Move into reach.

1. From the current head image, select a visible traversable floor point in the

frontal approach region of the target panel rather than far off to either side. Require the complete direct path and final chassis footprint to remain clear, with enough visible space for the later rotation and opening motion. If the current view is strongly oblique, prefer a farther supported floor point that increases standoff from the panel. If no such floor point is safely supported, do not force this approach.

1. Call move_chassis_to_floor_point with that current image and floor point.

The move preserves yaw, so reaching the floor point alone does not prove that the robot faces the panel. Inspect its returned post-action head image; never continue from the pre-motion image.

1. If the new view is still strongly oblique, first increase standoff from the

panel when the chassis is too close for a clear turn, then use a rotation-only adjust_chassis call: set spin and omit forward and translation. Choose its sign and magnitude from the current panel geometry, the clear rotation sweep, and measured tool feedback. Inspect each returned image. When the top and bottom edges of the panel look level, the base-forward direction is approximately normal to the target panel plane; continue to the next step. Do not use a fixed yaw target, spin amount, or retry count.

1. State the facing judgment from the latest post-rotation image. When that

judgment is facing, reidentify the target panel, hinge when applicable, and handle in that image. Only then continue to Move into reach from that current observation; never reuse pre-staging or pre-rotation coordinates. Complete Set trunk height after facing before any move_to_reach_point.

Set trunk height after facing

Complete this section after the latest stated judgment is facing and before any move_to_reach_point on this panel. Classify the target from the current image, not from task memory.

• Floor-standing closures: a cabinet, TV stand, washer, or similar body that

sits on the floor, whose door or drawer is at lower-body height. Call adjust_height with a negative upward large enough to reach the lowest reachable body height. Out-of-range requests clamp to that lowest row. If the trunk is already at that lowest reachable height, skip the call.

• High closures: a wall cabinet, overhead cabinet, or other door whose

handle sits well above standing torso height. Call reset_body with keep_ori_arm="none" (or omit that argument). reset_body is a trunk/posture tool, not an episode reset. A parent prompt such as "Do not reset" forbids resetting the episode; it does not forbid this posture. If the trunk is already upright, skip the call.

Inspect the returned head image. Re-state the facing judgment from that image. When it is not facing, return to Establish a frontal staging pose. When it is facing, continue to Move into reach from that current observation; never reuse pre-height coordinates.

Move into reach

This section is for a pose whose latest stated judgment is facing. Do not call move_to_reach_point until Set trunk height after facing has been completed for this panel.

1. State the facing judgment from the current head image, with a short

panel-plane note. When that judgment is not facing, return to Establish a frontal staging pose and finish that section first.

1. In the current head image, select a visible point on the closure surface

immediately above its handle. The point must lie on a depth-bearing physical surface, not in open air, on background beyond the closure, or on the handle itself.

1. Call move_to_reach_point with that current image_id, u, and v only

when the current image shows a clear straight path to that point. Omit reach, nav_timeout_s, and keep_ori_arm unless current evidence requires a non-default value.

1. Inspect shoulder_distance_estimate in the completed call before grasp

planning. Treat the pose as reach-ready only when at least one finite left_m or right_m value is strictly less than 0.7 m. This is a distance gate, not an instruction to choose that arm. A passing distance gate still requires a facing judgment from the post-action image. After this move, state facing again from that image using the panel-plane look.

1. If neither shoulder passes the gate, or the current post-action image shows

an awkward approach or pull pose, reposition before planning. When that image still shows the panel plane at a large inclination, return to Establish a frontal staging pose, get farther from the panel, rotate until the panel plane faces the robot, and then call move_to_reach_point. When the panel plane already faces the robot, from current geometry, clearance, and tool feedback, either rerun move_to_reach_point on a newly selected valid closure-surface point or use adjust_chassis for an appropriate forward, backward, or lateral correction. Choose the direction and magnitude from the scene; do not use a fixed correction, retry sequence, or arbitrary exact shoulder-distance target.

1. Inspect every repositioning result and use its post-action head image when

usable. After an adjust_chassis correction, do not reuse the old image or shoulder estimate; use the returned current image to obtain a fresh move_to_reach_point result before planning. Once the current move result passes the distance gate and the latest facing judgment is facing, reidentify the closure, handle, and, for a hinged closure, the hinge side. Never plan from a pre-motion image. When the current image is a long low TV stand drawer and that current move result already passes the distance gate, go to TV stand drawer special case and stay there. Do not continue to Plan and grasp the handle or Probe the grasp and open the hinged door.

Ground the physical handle

A long low TV stand drawer that already passed reach does not use this section. Go to TV stand drawer special case. This eight-point bind is for the hinged-door path.

Before selecting any grasp point:

1. Classify the target closure mechanism from current visual evidence before

evaluating handle candidates:

• A hinged door rotates about a visible or geometrically supported hinge

axis. Its handle may be a vertical bar, horizontal pull, recessed pull, or knob on either side.

• A drawer translates outward and has no hinge. Its pull or knob may be

centered or offset on the drawer front.

• A sliding or glass door translates along a track. It may have a recessed

pull, projecting handle, or frame-integrated grip; a panel stile alone is not a handle. If the mechanism is ambiguous, obtain a clearer current observation instead of using handle orientation or position to guess the mechanism.

1. Identify the target movable panel and trace its visible boundary. A valid

handle candidate must belong to that panel, not to a fixed frame, control panel, cabinet face, or neighboring closure.

1. Accept a candidate only with positive visual evidence of a bounded

manipulable component attached to that panel: a projecting bar or knob, a bounded recessed cavity or lip with visible depth-bearing grasp structure, or another finite clampable body. Bounded extent, distinct surface geometry, local attachment, clearance, or depth separation may support identity; no single cue is mandatory. Color or expected layout alone is insufficient.

1. Use no universal orientation or side prior. Determine each candidate's

orientation, visible extent, and attachment from the current image. A valid candidate may be vertical, horizontal, recessed, frame-integrated, a knob, or near any side.

1. Reject panel borders, seams, decorative trim, continuous frame stiles,

glass, reflections, hinges, buttons, controls, and other closures' handles; do not grasp attachment endpoints or corners. Treat a boundary-following strip as a seam, lip, trim, or stile only when current evidence supports it as continuous structure without a bounded manipulable body. Boundary alignment or missing visible clearance alone is not enough to accept or reject it.

1. Treat the selected point as the planner's intended gripper-center target,

not merely a semantic label for the handle. The camera ray through that point must terminate on solid, graspable handle material that the closed gripper can clamp. Never select a handle opening, through-hole, empty gap, or other negative space, including background visible through the handle. Do not select the floor or back wall of a recessed cavity unless that surface is itself the solid handle body to be clamped. For a recessed or loop handle, select a reachable solid lip, bar, or clamping surface rather than the empty center. Select near the center of that exposed solid surface, away from endpoints and the door-frame seam. If the handle body remains indistinguishable from these confusers or no solid grasp point is visible, obtain a clearer observation; never guess or plan a nearby point. Stay in this Skill.

1. When a recessed or loop handle borders negative space and a single RGB point

cannot be separated confidently from its aperture, validate candidates with measured depth before grasp planning. This procedure applies only after the bounded component has positive visual evidence of being a handle. It is forbidden for a circular appliance-door rim, aperture, bezel, or panel; use the Circular appliance door special case instead. First inspect persistent_tracking: track_object_distance is an atomic full-set replacement. Do not destroy active bindings required by the parent task; if such bindings exist and RGB evidence is insufficient, do not replace those bindings. Obtain a clearer observation instead. Otherwise, in one track_object_distance call bind at most eight points on the same current image: up to six distinct solid-handle candidates distributed along the bounded lip, ledge, or bar; one deliberate aperture/recess-interior reference; and one broad movable-panel reference. The two reference points are measurement controls and will never be used as a grasp target. Do not cluster all candidates on one ambiguous patch or reuse the aperture center as a nominal handle candidate.

Reject any nominal handle candidate whose measured depth clusters with the aperture/recess reference or lies at a depth discontinuity into that negative space. Depth never proves handle identity: a surviving point must still be visibly inside the same bounded, clampable handle component rather than the panel, ring, frame, seam, or trim. Prefer a central candidate with visible material margin. Re-locate that candidate in the latest clickable head image after the measurement, then pass its current image_id, u, and v relative-coordinate tuple to grasp planning. Do not reuse the pre-measurement image tuple when tracking has returned a newer image. If no candidate passes both visual identity and measured-depth checks, obtain a clearer observation and ground again. Do not rebind repeatedly or scatter new measured points after a failed set. Stay in this Skill.

1. Use integer relative coordinates in the latest clickable head image:

u and v range from 0 through 1000. Use the full image margins, never coordinates relative to the appliance, target panel, handle, or a crop. Treat image_id, u, and v as one inseparable observation tuple. Pass the exact selected image_id; never combine coordinates from different images.

1. If the handle remains indistinguishable from the frame, obtain a clearer

current observation and ground it again. Do not probe nearby planner points. After failure, do not scatter new points along the same presumed feature without first re-verifying it as the handle.

Microwave special case

Apply this section only after the current image independently establishes this microwave layout: a windowed movable door on the left and a fixed control panel on the right. Other microwave layouts use only the general grounding rules above.

1. Treat the narrow, bounded vertical bar attached to the door's right edge,

immediately left of the fixed control panel, as the handle candidate. It must still satisfy movable-panel membership and positive grasp evidence.

1. Use finite top and bottom extent, its own visible body thickness or shading,

and localized attachment to the movable door as positive evidence. One RGB view does not need to expose an open-air gap behind the bar. Do not reject that resolved bar solely because its rear gap is not visible.

1. Reject the pale control panel, circular controls, red buttons, straight

door/control seam, window frame, reflections, and robot parts. The long dark horizontal strip along the window is trim, not the handle.

1. When unobstructed, select inside the middle third of the visible bar body,

away from its endpoints and the adjacent seam. If an arm or gripper hides any of that central region, obtain a new observation instead of inferring a point.

Never transfer these asset-specific cues to other doors, drawers, or microwave layouts. Without the verified entry layout, do not infer that a handle is vertical or on the right.

Circular appliance door special case

Apply this section when the current image shows a front-loading washing-machine or dryer porthole door: one large circular door or window set into the appliance front. The circular shape alone does not identify which ring surface moves with the door.

1. Do not classify the circular rim, window aperture, bezel, or surrounding

panel as a recessed or loop handle merely because they form concentric depth layers. Do not call track_object_distance on rim/aperture/panel candidates for this mechanism. Those measurements separate surfaces by depth but cannot prove movable-door membership, an undercut, or gripper clearance.

1. Require positive current-image evidence of a finite handle or latch attached

to the movable door: for example a projecting tab or bar, or a bounded solid lip with a visible undercut and enough material for the closed gripper to clamp. Determine its location from the image; do not assume a side from the appliance category. The ring itself is not a handle by default.

1. After the handle is identified, follow Plan and grasp the handle and

Probe the grasp and open the hinged door. An empty-looking close or missing bilateral contact is not a reason to skip the EEF probe. If the arm hides the mechanism before planning, move that arm to its grasp-ready observation pose once; if the viewpoint still cannot establish a finite handle, make one clearance-supported viewpoint correction and use its new head image. Re-ground a different solid component only from that changed observation.

1. If no positively identified clampable handle or latch is visible, continue

changing the observation rather than planning on the circular ring or using depth tracking as a substitute for handle identity. Stay in this Skill.

TV stand drawer special case

<!-- 电视柜抽屉：独立分支，不进铰链门主路径 --> This is a separate branch, not a variant of the hinged-door path. After Set trunk height after facing and Move into reach, a long low TV stand drawer stays in this section until the drawer is open or this Skill exits. Do not enter Plan and grasp the handle, Retract before a new grasp plan, Probe the grasp and open the hinged door, or Clear around an opened door before the next.

Apply this section only when the current image shows a long low TV stand (media console) pull-out drawer, not a coffee table, wall cabinet, washer, microwave, or hinged door. Open only the requested drawer. Do not open the other drawer or the glass door. Do not close any drawer.

Do not call plan_grasp_point_filter_rgbd_lite on this mechanism. Do not call wrist-frame adjust_*_eef_pose_in_wrist_frame with forward=-0.02. After a passing shoulder_distance_estimate (left_m or right_m strictly below 0.7 m) from the current move_to_reach_point, do not call measure_shoulder_distance and do not treat a chassis-forward warning on the TV stand face as a reason to reverse or to skip the next step. Do not use the eight-point recessed-handle bind in Ground the physical handle.

<!-- 电视柜把手：柜面最上沿那条横向细条，不是柜面黑斑或旋钮 --> The handle is the thin horizontal strip along the top edge of the requested drawer front, just under the stand's top overhang. It runs the width of that drawer as one long solid bar.

Then do these in order.

1. In the current head image, call track_object_distance once on the

solid bar body near the center of that top strip. Name that point handle.

1. Call move_tracked_point with execution_mode=plan and this input:

C1, vertical pinch through handle (same x,y): left_finger_tip on_hand x,y,a handle off_hand x,y,z right_finger_tip on_hand x,y,b C2, gripper slide at the same height as handle (same z): handle off_hand x,y,z gripper_slide_center on_hand c,d,z Use this input as written. If the plan fails or the red overlay misses handle, recapture, bind handle again, and call again with this same input.

1. Call exec_plan_pose with that plan_id and back_m=0.1. If this

call fails, open_gripper if that gripper is closed, then set_arm_to_grasp_position for that arm with gripper="open" (or arm="both" if the other arm is also out), then return to step 2 with the same input.

1. Call close_gripper on that same arm.
2. Call adjust_chassis with forward=-0.3 and look at the returned head

image. If the drawer is open: open_gripper on that arm, then set_arm_to_grasp_position with arm="both" and gripper="open", then leave this Skill. If the drawer is still closed, open_gripper if needed, then set_arm_to_grasp_position with arm="both" and gripper="open", then repeat from step 1 with the same input.

Plan and grasp the handle

This section is the hinged-door path. A long low TV stand drawer must not enter this section. Use TV stand drawer special case instead.

1. After the latest facing judgment is facing, move_to_reach_point has

passed the shoulder gate, and the handle is grounded, bind one pull-witness point on solid handle material with track_object_distance on the current image. Name that point so it can be reread after each pull. This bind is the motion witness, not a second grasp target; planning still uses the grounded handle body, which may be the same point. track_object_distance is an atomic full-set replacement. Do not destroy active bindings required by the parent task; if such bindings exist, obtain a clearer observation instead of replacing them. On a circular appliance door, do not bind rim, aperture, bezel, or panel points; only a positively identified handle may be bound.

1. After completing the handle-grounding checks above, select a visible grasp

point on the handle in the latest usable head image and call plan_grasp_point_filter_rgbd_lite with that exact image and point. Always omit the plan_arm argument and let the planner select whichever arm has a collision-safe reachable grasp. Do not constrain the grasp-planning arm from the hinge side, handle side, closure type, or a prior arm assumption. The hinge side controls the later pull direction only, never the grasp-planning arm. Do not pass the literal default value either.

1. Inspect the returned plan and grasp preview. Continue only when the plan is

successful and the preview places the selected gripper on the intended handle without an evident collision.

1. Call exec_plan_pose with the returned plan_id. This tool moves the EEF

but does not close the gripper.

1. After exec_plan_pose, inspect the returned wrist image. Do not call a

wrist camera just to replace that image. Red overlay points are visible non-robot 3D points inside that gripper's grasp volume, between the two fingers. Judge whether the grasp pose is good enough to close from that overlay: both fingers should straddle only the movable handle, and judge whether handle material is visibly between the two fingers. The red must overlap the handle alone. The adjacent sidewall, fixed cabinet, frame, seam, or a neighboring closure must not have red; if they do, closing will clamp the handle and that fixed surface together and the door or drawer cannot open. Neither finger may rest on the fixed cabinet, frame, sidewall, seam, or a neighboring closure. Record that look; it does not decide whether to pull. Do not close yet.

1. After that inspect, first call open_gripper once on the same arm so

the fingers are fully open. Then call that arm's adjust_left_eef_pose_in_wrist_frame or adjust_right_eef_pose_in_wrist_frame with only forward=-0.02. Inspect the new wrist overlay and judge whether the red left the sidewall or otherwise changed. Then call that same tool once more with only forward=-0.02 and inspect the newest wrist image. Do not use forward=-0.01. Do not use a positive forward to seat deeper. Inspect again and close only when the red is on the handle alone. If the red still covers the handle and the sidewall after those two 2 cm retracts, do not close; Translation may use forward, leftward, and upward (positive or negative). Prefer translation-only calls first. If rotation is needed, change only one of roll, pitch, or yaw by 10 or -10 degrees per call; do not use a larger rotation step. Rotation moves the wrist-frame axes, so it also changes later translations: inspect the new wrist image after each rotation before translating again, and do not combine a rotation with a translation in the same call. If several later adjusts still cannot isolate the handle, do not close; follow Retract before a new grasp plan, then return to Ground the physical handle and replan. Do not skip the open_gripper or the two forward=-0.02 retracts after exec_plan_pose.

1. Use the arm selected by the successful plan as the holding arm. Call

close_gripper for exactly that arm. Missing bilateral contact, an empty-looking close, or an uncertain wrist view is not a reason to skip the pull probe below. Stay in this Skill.

Retract before a new grasp plan

Use this section when an already-executed grasp is abandoned and a new plan_grasp_point_filter_rgbd_lite is needed: the wrist pose cannot be corrected, the close missed, the probe did not move the handle, or the chosen arm or point is being replaced. Replaying the current plan_id after open_gripper is not this section.

1. If that gripper is closed, call open_gripper on the unsuccessful arm.
2. Call set_arm_to_grasp_position for that unsuccessful arm with

gripper="open" so it leaves the handle. If the other arm is also extended near the same handle, panel, or a neighboring point, call set_arm_to_grasp_position with arm="both" and gripper="open" instead. Do not leave an arm parked at the target.

1. Only after that returned head image, re-ground the handle and call

plan_grasp_point_filter_rgbd_lite. Never plan or exec_plan_pose for a second arm onto a nearby point while the first arm is still at the handle or panel.

1. Only one arm may occupy the handle region at a time.

Probe the grasp and open the hinged door

The grasp test is whether the tracked handle point moves with a handle-opening EEF pull. Close telemetry does not decide this.

1. First probe without moving the chassis. Call the holding arm's

adjust_left_eef_pose_in_head_frame or adjust_right_eef_pose_in_head_frame with a pull away from the panel toward the robot. When the robot is facing the panel, use about forward=-0.2 in the head-camera frame. Omit chassis motion on this probe. Do not rotate the chassis.

1. Read the tracked handle point. If its measured position moved in the same

sense as that EEF pull, by a comparable amount rather than noise, the handle is held. Then open the door by coordinating further same-arm EEF pulls with adjust_chassis. Call adjust_chassis with a bounded combined robot-frame translation and omit spin:

• Left-side hinge: use forward=-0.1 and translation=0.1 meters.
• Right-side hinge: use forward=-0.1 and translation=-0.1 meters.

The tool combines these two components into one same-frame two-dimensional chassis translation. Do not exceed 0.1 m in magnitude on either chassis axis and never rotate the chassis while the handle is grasped. After each coordinated step, confirm the handle point still moved with the motion. If the full chassis step would no longer follow the observed door arc, use a smaller forward and/or translation magnitude on the next adjust_chassis call while preserving the hinge-specific signs above. Never reverse either sign merely to force progress. A first successful probe or a small opening is not completion. Keep repeating coordinated same-arm EEF pulls and adjust_chassis steps along the door arc. Those two tools together can carry the door through the full swing; do not leave the handle after the panel has only started to move.

1. If the handle point did not move with the probe, the gripper did not take

the handle. Call open_gripper on that arm, then exec_plan_pose again on the current plan, then repeat the same post-exec sequence: call open_gripper once, then forward=-0.02, inspect the wrist image, then forward=-0.02 again and inspect, then call close_gripper twice in a row on that same arm. Then repeat the chassis-still EEF probe and again judge by handle-point motion. Stay in this Skill.

1. Leave this Skill only after a current observation shows that the

hinged closure is fully open, as close to 90 degrees as the swing allows. Report success without commanding any closing motion. Do not open the gripper or otherwise release the handle just because the door started to move. When that full-open observation is current, end this opening: if a gripper still holds the handle, call open_gripper on that arm, then call set_arm_to_grasp_position with arm="both" and gripper="open". Then leave this Skill. If another door or leaf still needs opening, stay in this Skill and continue with Clear around an opened door before the next.

Clear around an opened door before the next

Use this section after the current hinged door is fully open and another door or leaf still needs opening, or when this Skill is entered while an already-open door occupies the approach to the next target.

1. If a gripper still holds the opened door, call open_gripper on that

arm, then continue with the retreat.

1. Call adjust_chassis with forward=-0.5 meters and omit spin and

translation. This is a straight retreat away from the opened door.

1. Call set_arm_to_grasp_position with arm="both" and gripper="open".
2. In the current head image, select a visible traversable floor point

that routes around the opened door so the complete direct path and final chassis footprint remain clear of that open leaf and its swing.

1. Call move_chassis_to_floor_point with that current image and floor

point. Inspect the returned post-action head image and continue from that current image.

1. If the opened door or any other obstacle still blocks the straight

corridor between the robot and the next target panel, use adjust_chassis until that corridor is clear. Choose the direction and magnitude from the scene.

1. Then restart this Skill from Establish a frontal staging pose for

the next door. Re-state facing on that next panel. Use only coordinates from the latest post-clearance image.

Failures

• A long low TV stand drawer stays in TV stand drawer special case.

If exec_plan_pose fails or a pull leaves the drawer closed, first set_arm_to_grasp_position, then call again with the same input.

• On an explicit unsafe result, halt that pull or plan, reobserve, and

continue this Skill. Do not end the session.

• When a motion reports success but omits completion details, do not treat that

omission as completion or failure. Inspect a usable returned image or call capture_head_camera, then continue from the observed state.

• If the EEF probe or a later coordinated step produces no matching

handle-point motion, use the miss path above: open the gripper, re-exec the current plan, then open_gripper once and two forward=-0.02 wrist retracts with an inspect after each, close twice, and probe again. Do not skip the probe because the close looked uncertain. If that same-plan retry is abandoned for a new point or a different arm, follow Retract before a new grasp plan before the next plan_grasp_point_filter_rgbd_lite.

• Never leave one arm at the handle while the other plans or executes a

nearby grasp. Retract the unsuccessful arm with set_arm_to_grasp_position first. Only one arm may occupy the handle region at a time.

• If the door has opened only a little, that is not done. Keep coordinating

EEF and chassis along the observed arc until the panel is fully open, as close to 90 degrees as the swing allows. Do not release the handle after a small opening.

• After one door is fully open and another still needs opening, continue

with Clear around an opened door before the next: retreat, stow both arms, drive around the opened door, then restart staging.

• If grasp planning fails or reports the handle as unreachable before contact,

first reassess current shoulder distances and approach geometry. When the panel is still seen at a large inclination, recover facing from the panel plane first, then resume reach. When the evidence indicates a reach or pose problem while the panel plane already faces the robot, return to Move into reach and reposition instead of scattering nearby grasp points or repeating the same plan from the same pose.